By assuming that the entropy of an isolated system changes so as to yield maximum probability all through its transient, and by respecting the fact that Onsager's regression hypothesis originally refers to conditional ensemble averages, we present here the notion that the regression of the variables which define the system is a manifestation of the regression of the entropy of the system in the multi-dimensional space composed of those variable coordinates. Due to our new interpretation of Onsager Casimir symmetry, no upper limit of correlation time exists for its validity.